Home

Awesome

RangeLoss-Caffe

My implement of <Range Loss for Deep Face Recognition with Long-tail>

Note

If you want to train with range loss, please add:

message RangeLossParameter {
  optional int32 choose_k = 1 [default = 2];
  optional float inter_weight = 2 [default = 10e-4];
  optional float intra_weight = 3 [default = 10e-5];
  optional float margin = 4 [default = 2e4];
}

to the src/caffe/proto/caffe.proto

and:

  optional RangeLossParameter range_loss_param = 155;

to the end of "message LayerParameter"

The UniformLayer is used to get the uniform training data from filelist

please add:

optional uint32 uniform_num = 14 [default = 1];

to the end of "message ImageDataParameter"

Note

The implementation is a little different from the original formula in the paper, I will check them again, welcome to dissucess.

train with softmaxloss on minist